Issue 73: Smooth using custom prior configs and a new vignette

.github/copilot-instructions.md

@@ -15,7 +15,7 @@ NowcastAutoGP.jl is a **Julia package for forecasting with nowcasting** that com
 
 ### Main Pipeline (see `docs/vignettes/tutorial.qmd` for complete example)
 1. **Transform Data**: `create_transformed_data(dates, values; transformation)` using `get_transformations()`
-2. **Fit Base Model**: `make_and_fit_model(data)` on historical "confirmed" data
+2. **Fit Base Model**: `make_and_fit_model(data; n_mcmc=200, n_hmc=50)` on historical "confirmed" data (`n_mcmc`/`n_hmc` are required pass-through kwargs of `AutoGP.fit_smc!`)
 3. **Create Nowcasts**: Generate multiple `TData` objects representing uncertain recent data scenarios
 4. **Forecast with Uncertainty**: `forecast_with_nowcasts()` temporarily adds each nowcast to model, generates forecasts, then removes it
 
@@ -40,14 +40,15 @@ NowcastAutoGP.jl is a **Julia package for forecasting with nowcasting** that com
 ```julia
 # Standard pattern for fitting
 model = make_and_fit_model(data;
-    n_particles=8,           # SMC particles
-    smc_data_proportion=0.1, # Data proportion per SMC step
-    n_mcmc=200,             # MCMC structure steps
-    n_hmc=50                # HMC parameter steps
+    n_particles=8,           # SMC particles (forwarded to AutoGP.GPModel)
+    smc_data_proportion=0.1, # Data proportion per SMC step (wrapper-only; builds the schedule)
+    config=GPConfig(),       # GP prior; pass a customised GPConfig() to change kernel structure / priors
+    n_mcmc=200,              # forwarded to AutoGP.fit_smc! via kwargs...
+    n_hmc=50                 # forwarded to AutoGP.fit_smc! via kwargs...
 )
 ```
 
-**Key Implementation Detail**: `effective_proportion = max(smc_data_proportion, 1.0/n_train)` prevents schedule step=0 errors with small datasets.
+**Key Implementation Detail**: Only `n_particles`, `smc_data_proportion`, and `config` are named wrapper kwargs; everything else (`n_mcmc`, `n_hmc`, `hmc_config`, `biased`, `shuffle`, `verbose`, …) is passed straight through `kwargs...` to `AutoGP.fit_smc!` rather than re-declared here. `effective_proportion = max(smc_data_proportion, 1.0/n_train)` prevents schedule step=0 errors with small datasets.
 
 ## Forecasting with Nowcasts (`src/forecasting.jl`)
 

docs/Project.toml

@@ -1,4 +1,5 @@
 [deps]
+Accessors = "7d9f7c33-5ae7-4f3b-8dc6-eff91059b697"
 AutoGP = "6eb593e7-dfb4-4e48-b98e-d7222cdf0053"
 CSV = "336ed68f-0bac-5ca0-87d4-7b16caf5d00b"
 CairoMakie = "13f3f980-e62b-5c42-98c6-ff1f3baf88f0"

docs/make.jl

@@ -7,6 +7,7 @@ pages = [
     "Getting Started" => "vignettes/getting-started.md",
     "Tutorials" => [
         "vignettes/use-percentages.md",
+        "vignettes/setting-priors.md",
     ],
     "API Reference" => "api.md",
 ]

docs/make_vignettes.jl

@@ -1,5 +1,9 @@
 using Literate
-vignettes_to_make = ["getting-started.jl", "use-percentages.jl"]
+vignettes_to_make = [
+    "getting-started.jl",
+    "use-percentages.jl",
+    "setting-priors.jl",
+]
 
 # Generate markdown with executed output from Literate.jl scripts.
 # Uses CommonMarkFlavor so Documenter.jl won't re-evaluate code blocks.